0% found this document useful (0 votes)
369 views64 pages

Onur Digitaldesign 2020 Lecture19 Simd Beforelecture

COA lectures by prof onur mutlu ETH Zurich

Uploaded by

Manish sutradhar
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
369 views64 pages

Onur Digitaldesign 2020 Lecture19 Simd Beforelecture

COA lectures by prof onur mutlu ETH Zurich

Uploaded by

Manish sutradhar
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 64

Digital Design & Computer Arch.

Lecture 19: SIMD Processors

Prof. Onur Mutlu

ETH Zürich
Spring 2020
7 May 2020
We Are Almost Done With This…
◼ Single-cycle Microarchitectures

◼ Multi-cycle Microarchitectures

◼ Pipelining

◼ Issues in Pipelining: Control & Data Dependence Handling,


State Maintenance and Recovery, …

◼ Out-of-Order Execution

◼ Other Execution Paradigms


2
Approaches to (Instruction-Level) Concurrency
◼ Pipelining
◼ Out-of-order execution
◼ Dataflow (at the ISA level)
◼ Superscalar Execution
◼ VLIW
◼ Systolic Arrays
◼ Decoupled Access Execute
◼ Fine-Grained Multithreading
◼ SIMD Processing (Vector and array processors, GPUs)

3
Readings for this Week
◼ Required
◼ Lindholm et al., "NVIDIA Tesla: A Unified Graphics and
Computing Architecture," IEEE Micro 2008.

◼ Recommended
❑ Peleg and Weiser, “MMX Technology Extension to the Intel
Architecture,” IEEE Micro 1996.

4
Announcement
◼ Late submission of lab reports in Moodle
❑ Open until June 20, 2020, 11:59pm (cutoff date -- hard
deadline)
❑ You can submit any past lab report, which you have not
submitted before its deadline
❑ It is NOT allowed to re-submit anything (lab reports, extra
assignments, etc.) that you had already submitted via other
Moodle assignments
❑ We will grade your reports, but late submission has a
penalization of 1 point, that is, the highest possible score per
lab report will be 2 points

5
Exploiting Data Parallelism:
SIMD Processors and GPUs
SIMD Processing:
Exploiting Regular (Data) Parallelism
Flynn’s Taxonomy of Computers
◼ Mike Flynn, “Very High-Speed Computing Systems,” Proc.
of IEEE, 1966

◼ SISD: Single instruction operates on single data element


◼ SIMD: Single instruction operates on multiple data elements
❑ Array processor
❑ Vector processor
◼ MISD: Multiple instructions operate on single data element
❑ Closest form: systolic array processor, streaming processor
◼ MIMD: Multiple instructions operate on multiple data
elements (multiple instruction streams)
❑ Multiprocessor
❑ Multithreaded processor
8
Data Parallelism
◼ Concurrency arises from performing the same operation on
different pieces of data
❑ Single instruction multiple data (SIMD)
❑ E.g., dot product of two vectors

◼ Contrast with data flow


❑ Concurrency arises from executing different operations in parallel (in
a data driven manner)

◼ Contrast with thread (“control”) parallelism


❑ Concurrency arises from executing different threads of control in
parallel

◼ SIMD exploits operation-level parallelism on different data


❑ Same operation concurrently applied to different pieces of data
❑ A form of ILP where instruction happens to be the same across data
9
SIMD Processing
◼ Single instruction operates on multiple data elements
❑ In time or in space
◼ Multiple processing elements

◼ Time-space duality

❑ Array processor: Instruction operates on multiple data


elements at the same time using different spaces

❑ Vector processor: Instruction operates on multiple data


elements in consecutive time steps using the same space

10
Array vs. Vector Processors
ARRAY PROCESSOR VECTOR PROCESSOR

Instruction Stream Same op @ same time


Different ops @ time
LD VR  A[3:0] LD0 LD1 LD2 LD3 LD0
ADD VR  VR, 1 AD0 AD1 AD2 AD3 LD1 AD0
MUL VR  VR, 2
ST A[3:0]  VR MU0 MU1 MU2 MU3 LD2 AD1 MU0
ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0
Different ops @ same space AD3 MU2 ST1
MU3 ST2
Time Same op @ space ST3

Space Space

11
SIMD Array Processing vs. VLIW
◼ VLIW: Multiple independent operations packed together by the compiler

12
SIMD Array Processing vs. VLIW
◼ Array processor: Single operation on multiple (different) data elements

13
Vector Processors (I)
◼ A vector is a one-dimensional array of numbers
◼ Many scientific/commercial programs use vectors
for (i = 0; i<=49; i++)
C[i] = (A[i] + B[i]) / 2

◼ A vector processor is one whose instructions operate on


vectors rather than scalar (single data) values
◼ Basic requirements
❑ Need to load/store vectors → vector registers (contain vectors)
❑ Need to operate on vectors of different lengths → vector length
register (VLEN)
❑ Elements of a vector might be stored apart from each other in
memory → vector stride register (VSTR)
◼ Stride: distance in memory between two elements of a vector

14
Vector Processors (II)
◼ A vector instruction performs an operation on each element
in consecutive cycles
❑ Vector functional units are pipelined
❑ Each pipeline stage operates on a different data element

◼ Vector instructions allow deeper pipelines


❑ No intra-vector dependencies → no hardware interlocking
needed within a vector
❑ No control flow within a vector
❑ Known stride allows easy address calculation for all vector
elements
◼ Enables prefetching of vectors into registers/cache/memory

15
Vector Processor Advantages
+ No dependencies within a vector
❑ Pipelining & parallelization work really well
❑ Can have very deep pipelines, no dependencies!

+ Each instruction generates a lot of work


❑ Reduces instruction fetch bandwidth requirements

+ Highly regular memory access pattern

+ No need to explicitly code loops


❑ Fewer branches in the instruction sequence

16
Vector Processor Disadvantages
-- Works (only) if parallelism is regular (data/SIMD parallelism)
++ Vector operations
-- Very inefficient if parallelism is irregular
-- How about searching for a key in a linked list?

Fisher, “Very Long Instruction Word architectures and the ELI-512,” ISCA 1983. 17
Vector Processor Limitations
-- Memory (bandwidth) can easily become a bottleneck,
especially if
1. compute/memory operation balance is not maintained
2. data is not mapped appropriately to memory banks

18
Vector Processing in More Depth
Vector Registers
◼ Each vector data register holds N M-bit values
◼ Vector control registers: VLEN, VSTR, VMASK
◼ Maximum VLEN can be N
❑ Maximum number of elements stored in a vector register
◼ Vector Mask Register (VMASK)
❑ Indicates which elements of vector to operate on

❑ Set by vector test instructions

◼ e.g., VMASK[i] = (Vk[i] == 0)


M-bit wide M-bit wide
V0,0 V1,0
V0,1 V1,1

V0,N-1 V1,N-1

20
Vector Functional Units
◼ Use a deep pipeline to execute
element operations
V V V
→ fast clock cycle
1 2 3

◼ Control of deep pipeline is


simple because elements in
vector are independent

Six stage multiply pipeline

V1 * V2 → V3

Slide credit: Krste Asanovic 21


Vector Machine Organization (CRAY-1)
◼ CRAY-1
◼ Russell, “The CRAY-1
computer system,”
CACM 1978.

◼ Scalar and vector modes


◼ 8 64-element vector
registers
◼ 64 bits per element
◼ 16 memory banks
◼ 8 64-bit scalar registers
◼ 8 24-bit address registers

22
CRAY X-MP-28 @ ETH (CAB, E Floor)

23
CRAY X-MP System Organization

Cray Research Inc., “The


CRAY X-MP Series of
Computer Systems,” 1985
24
CRAY X-MP Design Detail

Cray Research Inc., “The


CRAY X-MP Series of
Computer Systems,” 1985
25
CRAY X-MP CPU Functional Units

Cray Research Inc., “The


CRAY X-MP Series of
Computer Systems,” 1985
26
CRAY X-MP System Configuration

Cray Research Inc., “The


CRAY X-MP Series of
Computer Systems,” 1985
27
Seymour Cray, the Father of Supercomputers

"If you were plowing a field, which would you


rather use: Two strong oxen or 1024 chickens?"

© amityrebecca / Pinterest. https://www.pinterest.ch/pin/473018767088408061/

© Scott Sinklier / Corbis. http://america.aljazeera.com/articles/2015/2/20/the-short-brutal-life-of-male-chickens.html

28
Vector Machine Organization (CRAY-1)
◼ CRAY-1
◼ Russell, “The CRAY-1
computer system,”
CACM 1978.

◼ Scalar and vector modes


◼ 8 64-element vector
registers
◼ 64 bits per element
◼ 16 memory banks
◼ 8 64-bit scalar registers
◼ 8 24-bit address registers

29
Loading/Storing Vectors from/to Memory
◼ Requires loading/storing multiple elements

◼ Elements separated from each other by a constant distance


(stride)
❑ Assume stride = 1 for now

◼ Elements can be loaded in consecutive cycles if we can


start the load of one element per cycle
❑ Can sustain a throughput of one element per cycle

◼ Question: How do we achieve this with a memory that


takes more than 1 cycle to access?
◼ Answer: Bank the memory; interleave the elements across
banks
30
Memory Banking
◼ Memory is divided into banks that can be accessed independently;
banks share address and data buses (to minimize pin cost)
◼ Can start and complete one bank access per cycle
◼ Can sustain N parallel accesses if all N go to different banks

Bank Bank Bank Bank


0 1 2 15

MDR MAR MDR MAR MDR MAR MDR MAR

Data bus

Address bus

CPU
Picture credit: Derek Chiou 31
Vector Memory System
◼ Next address = Previous address + Stride
◼ If (stride == 1) && (consecutive elements interleaved
across banks) && (number of banks >= bank latency), then
❑ we can sustain 1 element/cycle throughput
Base Stride
Vector Registers

Address
Generator +

0 1 2 3 4 5 6 7 8 9 A B C D E F
Memory Banks
Picture credit: Krste Asanovic 32
Scalar Code Example: Element-Wise Avg.
◼ For I = 0 to 49
❑ C[i] = (A[i] + B[i]) / 2

◼ Scalar code (instruction and its latency)


MOVI R0 = 50 1
MOVA R1 = A 1 304 dynamic instructions
MOVA R2 = B 1
MOVA R3 = C 1
X: LD R4 = MEM[R1++] 11 ;autoincrement addressing
LD R5 = MEM[R2++] 11
ADD R6 = R4 + R5 4
SHFR R7 = R6 >> 1 1
ST MEM[R3++] = R7 11
DECBNZ R0, X 2 ;decrement and branch if NZ
33
Scalar Code Execution Time (In Order)
◼ Scalar execution time on an in-order processor with 1 bank
❑ First two loads in the loop cannot be pipelined: 2*11 cycles
❑ 4 + 50*40 = 2004 cycles

◼ Scalar execution time on an in-order processor with 16


banks (word-interleaved: consecutive words are stored in
consecutive banks)
❑ First two loads in the loop can be pipelined
❑ 4 + 50*30 = 1504 cycles

◼ Why 16 banks?
❑ 11-cycle memory access latency
❑ Having 16 (>11) banks ensures there are enough banks to
overlap enough memory operations to cover memory latency
34
Vectorizable Loops
◼ A loop is vectorizable if each iteration is independent of any
other

◼ For I = 0 to 49
❑ C[i] = (A[i] + B[i]) / 2
◼ Vectorized loop (each instruction and its latency):
MOVI VLEN = 50 1
7 dynamic instructions
MOVI VSTR = 1 1
VLD V0 = A 11 + VLEN – 1
VLD V1 = B 11 + VLEN – 1
VADD V2 = V0 + V1 4 + VLEN – 1
VSHFR V3 = V2 >> 1 1 + VLEN – 1
VST C = V3 11 + VLEN – 1
35
Basic Vector Code Performance
◼ Assume no chaining (no vector data forwarding)
❑ i.e., output of a vector functional unit cannot be used as the
direct input of another
❑ The entire vector register needs to be ready before any
element of it can be used as part of another operation
◼ One memory port (one address generator)
◼ 16 memory banks (word-interleaved)

1 1 11 49 11 49 4 49 1 49 11 49

V0 = A[0..49] V1 = B[0..49] ADD SHIFT STORE

◼ 285 cycles
36
Vector Chaining
◼ Vector chaining: Data forwarding from one vector
functional unit to another

V V V V V
LV v1 1 2 3 4 5
MULV v3,v1,v2
ADDV v5, v3, v4

Chain Chain

Load
Unit
Mult. Add

Memory

Slide credit: Krste Asanovic 37


Vector Code Performance - Chaining
◼ Vector chaining: Data forwarding from one vector
functional unit to another
1 1 11 49 11 49

Strict assumption:
Each memory bank
4 49 has a single port
(memory bandwidth
bottleneck)
These two VLDs cannot be 1 49
pipelined. WHY?

11 49

VLD and VST cannot be


◼ 182 cycles pipelined. WHY?
38
Vector Code Performance – Multiple Memory Ports
◼ Chaining and 2 load ports, 1 store port in each bank
1 1 11 49

1 11 49

4 49

1 49

11 49
◼ 79 cycles
◼ 19X perf. improvement!
39
Questions (I)
◼ What if # data elements > # elements in a vector register?
❑ Idea: Break loops so that each iteration operates on #
elements in a vector register
◼ E.g., 527 data elements, 64-element VREGs
◼ 8 iterations where VLEN = 64
◼ 1 iteration where VLEN = 15 (need to change value of VLEN)
❑ Called vector stripmining

40
(Vector) Stripmining

Source: https://en.wikipedia.org/wiki/Surface_mining 41
Questions (II)
◼ What if vector data is not stored in a strided fashion in
memory? (irregular memory access to a vector)
❑ Idea: Use indirection to combine/pack elements into vector
registers
❑ Called scatter/gather operations

42
Gather/Scatter Operations

Want to vectorize loops with indirect accesses:


for (i=0; i<N; i++)
A[i] = B[i] + C[D[i]]

Indexed load instruction (Gather)


LV vD, rD # Load indices in D vector
LVI vC, rC, vD # Load indirect from rC base
LV vB, rB # Load B vector
ADDV.D vA,vB,vC # Do add
SV vA, rA # Store result

43
Gather/Scatter Operations
◼ Gather/scatter operations often implemented in hardware
to handle sparse vectors (matrices)
◼ Vector loads and stores use an index vector which is added
to the base register to generate the addresses

◼ Scatter example
Index Vector Data Vector (to Store) Stored Vector (in Memory)

0 3.14 Base+0 3.14


2 6.5 Base+1 X
6 71.2 Base+2 6.5
7 2.71 Base+3 X
Base+4 X
Base+5 X
Base+6 71.2
Base+7 2.71
44
Conditional Operations in a Loop
◼ What if some operations should not be executed on a vector
(based on a dynamically-determined condition)?
loop: for (i=0; i<N; i++)
if (a[i] != 0) then b[i]=a[i]*b[i]

◼ Idea: Masked operations


❑ VMASK register is a bit mask determining which data element
should not be acted upon
VLD V0 = A
VLD V1 = B
VMASK = (V0 != 0)
VMUL V1 = V0 * V1
VST B = V1
❑ This is predicated execution. Execution is predicated on mask bit.
45
Another Example with Masking
for (i = 0; i < 64; ++i)
if (a[i] >= b[i]) Steps to execute the loop in SIMD code
c[i] = a[i]
1. Compare A, B to get
else
VMASK
c[i] = b[i]
2. Masked store of A into C

A B VMASK 3. Complement VMASK


1 2 0
2 2 1 4. Masked store of B into C
3 2 1
4 10 0
-5 -4 0
0 -3 1
6 5 1
-7 -8 1

46
Masked Vector Instructions
Simple Implementation Density-Time Implementation
– execute all N operations, turn off – scan mask vector and only execute
result writeback according to mask elements with non-zero masks

M[7]=1 A[7] B[7] M[7]=1


M[6]=0 A[6] B[6] M[6]=0 A[7] B[7]
M[5]=1 A[5] B[5] M[5]=1
M[4]=1 A[4] B[4] M[4]=1
M[3]=0 A[3] B[3] M[3]=0 C[5]

M[2]=0 C[4]
M[1]=1
M[2]=0 C[2]
M[0]=0
M[1]=1 C[1] C[1]

Write data port

M[0]=0 C[0]
Which one is better?
Write Enable Write data port
Tradeoffs?
Slide credit: Krste Asanovic 47
Some Issues
◼ Stride and banking
❑ As long as they are relatively prime to each other and there
are enough banks to cover bank access latency, we can
sustain 1 element/cycle throughput

◼ Storage of a matrix
❑ Row major: Consecutive elements in a row are laid out
consecutively in memory
❑ Column major: Consecutive elements in a column are laid out
consecutively in memory
❑ You need to change the stride when accessing a row versus
column

48
Matrix Multiplication
◼ A and B, both in row-major order

A4x6 B6x10 → C4x10


Dot products of rows and columns
of A and B

◼ A: Load A0 into vector register V1


❑ Each time, increment address by one to access the next column
❑ Accesses have a stride of 1
◼ B: Load B0 into vector register V2 Different strides can lead
❑ Each time, increment address by 10 to bank conflicts
❑ Accesses have a stride of 10
How do we minimize them?
49
Minimizing Bank Conflicts
◼ More banks

◼ Better data layout to match the access pattern


❑ Is this always possible?

◼ Better mapping of address to bank


❑ E.g., randomized mapping
❑ Rau, “Pseudo-randomly interleaved memory,” ISCA 1991.

50
Array vs. Vector Processors, Revisited
◼ Array vs. vector processor distinction is a “purist’s”
distinction

◼ Most “modern” SIMD processors are a combination of both


❑ They exploit data parallelism in both time and space
❑ GPUs are a prime example we will cover in a bit more detail

51
Recall: Array vs. Vector Processors
ARRAY PROCESSOR VECTOR PROCESSOR

Instruction Stream Same op @ same time


Different ops @ time
LD VR  A[3:0] LD0 LD1 LD2 LD3 LD0
ADD VR  VR, 1 AD0 AD1 AD2 AD3 LD1 AD0
MUL VR  VR, 2
ST A[3:0]  VR MU0 MU1 MU2 MU3 LD2 AD1 MU0
ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0
Different ops @ same space AD3 MU2 ST1
MU3 ST2
Time Same op @ space ST3

Space Space

52
Vector Instruction Execution
VADD A,B → C

Execution using Execution using


one pipelined four pipelined
functional unit functional units

A[6] B[6] A[24] B[24] A[25] B[25] A[26] B[26] A[27] B[27]
A[5] B[5] A[20] B[20] A[21] B[21] A[22] B[22] A[23] B[23]
A[4] B[4] A[16] B[16] A[17] B[17] A[18] B[18] A[19] B[19]
A[3] B[3] A[12] B[12] A[13] B[13] A[14] B[14] A[15] B[15]

C[2] C[8] C[9] C[10] C[11]

C[1] C[4] C[5] C[6] C[7]

Time Time
C[0] C[0] C[1] C[2] C[3]
Space
Slide credit: Krste Asanovic 53
Vector Unit Structure
Functional Unit

Partitioned
Vector
Registers
Elements 0, Elements 1, Elements 2, Elements 3,
4, 8, … 5, 9, … 6, 10, … 7, 11, …

Lane

Memory Subsystem

Slide credit: Krste Asanovic 54


Vector Instruction Level Parallelism
Can overlap execution of multiple vector instructions
❑ Example machine has 32 elements per vector register and 8 lanes
❑ Completes 24 operations/cycle while issuing 1 vector instruction/cycle

Load Unit Multiply Unit Add Unit


load
mul
add
time
load
mul
add

Instruction
issue

Slide credit: Krste Asanovic 55


We did not cover the following slides.
They are for your preparation for the
next lecture.

56
Automatic Code Vectorization
for (i=0; i < N; i++)
C[i] = A[i] + B[i];
Scalar Sequential Code Vectorized Code

load load load

Iter. 1 load load load

add Time add add

store store store

load
Iter. Iter.
Iter. 2 load 1 2 Vector Instruction

add
Vectorization is a compile-time reordering of
operation sequencing
 requires extensive loop dependence analysis
store
Slide credit: Krste Asanovic 57
Vector/SIMD Processing Summary
◼ Vector/SIMD machines are good at exploiting regular data-
level parallelism
❑ Same operation performed on many data elements
❑ Improve performance, simplify design (no intra-vector
dependencies)

◼ Performance improvement limited by vectorizability of code


❑ Scalar operations limit vector machine performance
❑ Remember Amdahl’s Law
❑ CRAY-1 was the fastest SCALAR machine at its time!

◼ Many existing ISAs include (vector-like) SIMD operations


❑ Intel MMX/SSEn/AVX, PowerPC AltiVec, ARM Advanced SIMD
58
SIMD Operations in Modern ISAs
SIMD ISA Extensions
◼ Single Instruction Multiple Data (SIMD) extension
instructions
❑ Single instruction acts on multiple pieces of data at once
❑ Common application: graphics
❑ Perform short arithmetic operations (also called packed
arithmetic)
◼ For example: add four 8-bit numbers
◼ Must modify ALU to eliminate carries between 8-bit values
padd8 $s2, $s0, $s1
32 24 23 16 15 8 7 0 Bit position

a3 a2 a1 a0 $s0

+ b3 b2 b1 b0 $s1

a3 + b3 a2 + b2 a1 + b1 a0 + b0 $s2

60
Intel Pentium MMX Operations
◼ Idea: One instruction operates on multiple data elements
simultaneously
❑ À la array processing (yet much more limited)

❑ Designed with multimedia (graphics) operations in mind


No VLEN register
Opcode determines data type:
8 8-bit bytes
4 16-bit words
2 32-bit doublewords
1 64-bit quadword

Stride is always equal to 1.

Peleg and Weiser, “MMX Technology


Extension to the Intel Architecture,”
IEEE Micro, 1996.
61
MMX Example: Image Overlaying (I)
◼ Goal: Overlay the human in image 1 on top of the background in image 2

Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro, 1996. 62
MMX Example: Image Overlaying (II)
Y = Blossom image X = Woman’s image

Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro, 1996. 63
Digital Design & Computer Arch.
Lecture 19: SIMD Processors

Prof. Onur Mutlu

ETH Zürich
Spring 2020
7 May 2020

You might also like